METHODOLOGY AND SOFTWARE FOR SEARCHING FOR ANOMALIES IN TELEMETRY DATA OF A SOLAR POWER PLANT BASED ON AN ARTIFICIAL NEURAL NETWORK - AUTOENCODER

Objectives: Development of a new methodology and software for detecting anomalies in the operation of PV-modules based on an artificial neural network type autoencoder trained on telemetry data from a solar power plant. Methods: The methodology is based on statistical studies of deviations of measured from recovered neural network values of current and voltage of all PV-modules of the power plant. Additionally, a criterion for evaluating the presence of faults in the operation of the PV-module based on statistical studies is introduced. Results: Using the developed methodology and software for anomaly detection in telemetry data over six months of observations with different evaluation criteria, from 14 to 45 anomalies were detected in 33 PV-modules. All cases were analyzed for the causes of the anomalies in the operation of the PV-modules. Conclusion: It was established that using four standard deviations for the daily average measured values of current ∆I and voltage ∆U as a criterion for detecting anomalies allows identifying faulty PV-modules. Using three and two standard deviations as a criterion for detecting anomalies helps identify reduced efficiency in PV-module operation due to degradation, excessive shading, and other factors.